MONAI Label CVE-2026-133968: AI agents found and exploited a zero-day in 11 minutes
MONAI Label ≤ 0.8.5: unauthenticated path-traversal RCE (CVE-2026-133968) exploited in the wild by autonomous AI attack agents. No released fix yet.
Published by Zero Hunt, an autonomous AI red team on an on-premise appliance running private AI: automated penetration testing for networks and infrastructure, black-box or gray-box, with a human approving every step that matters.
On 20 November 2026, Amazon Threat Intelligence published an account of the thing defenders have been bracing for: an attacker's autonomous AI agents, with no human at the keyboard, found a previously unknown flaw in open-source medical-imaging software and used it to get root on a university server. The step from mapping the application to writing files on the box took about 11 minutes. The software is MONAI Label, the flaw is now CVE-2026-133968, and — the part that matters this week — there is no fixed release to install.
Developing story — first published 13:05 CET (12:05 UTC), 21 November 2026. Updated as the vendor and CISA publish more.
At a glance
| CVE | CVE-2026-133968 |
| Product / affected versions | MONAI Label through 0.8.5 · unauthenticated in the default setup (MONAI_LABEL_AUTH_ENABLE=False) |
| Fixed in | No released fix as of 2026-11-21 · latest release 0.8.5 is vulnerable · patch merged in PR #1915 (2026-10-19), slated for the next release |
| CVSS | 9.8 Critical · CVSS 3.1 AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H · scored by MITRE (CNA); NVD status Received, not yet scored |
| Exploited in the wild | Yes · since September 2026 (NVD record; Amazon Threat Intelligence) |
| CISA KEV | Not listed (catalog version 2026.11.20) |
| Public PoC | No weaponized exploit code located as of 2026-11-21 · the technique (path traversal via the upload filename parameter) is described publicly |
| Official advisory | GHSA-qc22-rmgc-vg94 |
What happened: AI agents found the MONAI Label zero-day on their own
The write-up comes from AWS CISO CJ Moses and describes a campaign run with Cairn, an open-source autonomous AI penetration-testing framework. Amazon assesses with moderate confidence that a Chinese-nexus actor operated it — the prompts were written in Simplified Chinese, script timestamps were set to China Standard Time, and the targets were universities in East Asia outside mainland China. Amazon got visibility on 10 September 2026 because the operator left a Cairn instance publicly reachable with no authentication; the campaign itself ran roughly 48 hours from 8 September.
What makes this report different from the usual "AI wrote some malware" story is the autonomy. The agents did not get handed an exploit. On one server they spent about 16 hours and 148 sequential actions trying to abuse Redis — replication abuse, sandbox escape, config writes, memory corruption — and failed at all of it. They pivoted to a Log4Shell attempt, an SSRF against a co-located Qdrant instance, and credential sprays against PostgreSQL and SSH using 80 real usernames scraped from the university's own public pages and papers. Only then did they reach the medical-imaging application: they identified its framework, mapped its API endpoints, found an upload endpoint with a filename parameter, and tested it for path traversal. It was vulnerable.
"In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server," Moses wrote — and the agents "kept adapting for nearly a full day without human direction." The single human contribution in the whole sequence was handing the agents a proxy that tunneled into a second university network running the same vulnerable code.
This is not a one-off research toy. The Register notes a separate Gambit Security report on a different Cairn operator who ran at least 105 attacks and compromised at least 27 companies — a Fortune 500 hospitality firm, a major US airline, an industrial-supplies distributor, an online retailer — exfiltrating payment-card records. The same framework is already in both espionage and criminal hands.
Am I affected? MONAI Label ≤ 0.8.5 in the default setup
MONAI Label is Project MONAI's annotation and active-learning server for medical imaging — the kind of thing a hospital research group or a university radiology lab stands up next to a GPU. In its default configuration it ships with authentication off (MONAI_LABEL_AUTH_ENABLE=False), so the REST API — including the write endpoints below — is open to anyone who can reach it. Every release up to and including 0.8.5 (the current latest) is affected.
How the path-traversal zero-day becomes root
The bug is textbook CWE-22. PUT /datastore/image takes a caller-controlled image query parameter and uses it as the image ID without validation; LocalDatastore.add_image() appends the extension, joins it to the datastore directory and writes the body with shutil.copy() — no containment check. A value like ../../../../tmp/x writes wherever the process can write. PUT /datastore/label has the same flaw through save_label().
Arbitrary file write is not yet code execution, and this is where the agents' trick is worth reading. They wrote a Python reverse-shell payload to sitecustomize.py — a file the Python interpreter auto-imports at startup — then called a training API endpoint that spawns a fresh Python process. The new process imported sitecustomize.py, the payload fired, and because the service was running as root in that deployment, the shell was root. No memory corruption, no ROP chain: just write a file to the right path and let the application run it for you.
Remediation
Am I affected? Check the running version (monailabel --version or pip show monailabel), whether the API is reachable from untrusted networks, and whether auth is enabled. A quick, non-destructive probe of the flaw is a benign traversal write to a scratch path — if a file lands outside the datastore directory, you are exposed.
Patch — there is no fixed release yet. As of 2026-11-21 the latest MONAI Label release is 0.8.5 and it is vulnerable. The fix ("confine datastore write paths") is merged in PR #1915 on the main branch but has not been cut into a tagged release. If you can run from source, deploy a build that includes that commit; otherwise you are on compensating controls until the project ships the next version.
Can't patch now? Compensating controls.
- Take the API off any untrusted network. A MONAI Label server has no business being internet-reachable.
- Enable authentication (
MONAI_LABEL_AUTH_ENABLE=True). This closes the unauthenticated path that was exploited in the wild, though a logged-inANNOTATORcan still reach the endpoint — it is mitigation, not a fix. - Front the service with a reverse proxy that rejects
../and encoded traversal in theimage/labelparameters and restricts the/datastore/routes. - Run the service as an unprivileged user in a container with a read-only root filesystem, so a file write cannot reach
sitecustomize.pyon a system path or land as root. This alone breaks the observed chain.
Hunt for compromise. Amazon's published indicators are behavioral, so hunt on behavior, not file hashes:
PUT /datastore/imageor/datastore/labelrequests whoseimage/labelparameter contains.., encoded traversal, or an absolute path (T1190).- A new or modified
sitecustomize.py, or any.pywritten outside the datastore directory, followed by a training-endpoint call (persistence via Python startup, T1546; execution via Python, T1059.006). - The MONAI Label process — or a child Python process — spawning a shell or opening an outbound connection, especially as root.
- Password-spray bursts against PostgreSQL and SSH preceding the above (T1110.003), and new
.ssh/authorized_keysor cron entries on the host.
Eradicate and verify. Assume anything the service user could read is exposed. Remove dropped files and any sitecustomize.py implant, revert added SSH keys and cron jobs, rotate every credential reachable from the host, and rebuild the box if root was obtained. Then confirm the fix: re-run the traversal write and verify the file is rejected or confined — a version string alone does not prove the path is closed.
What is not known yet
- NVD still lists the record as Received; MITRE's 9.8 stands, but NVD has not published its own analysis.
- The CVE is not in CISA KEV as of catalog 2026.11.20, despite in-the-wild exploitation since September — expect that to change.
- Amazon's own post described the target generically as a "medical imaging AI application"; the MONAI Label identification comes from the NVD record and the GitHub advisory, whose timing and endpoint match exactly.
- The underlying model(s) Cairn used, the full victim count beyond the two named universities, and whether the fix will ship as 0.8.6 or a later tag are all still open.
When the attacker runs at machine speed, so must the defender
The uncomfortable detail in this report is not the path-traversal bug — those are routine — but that the whole kill chain, from hours of failed Redis attempts to a working root shell on a different service, ran for nearly a day with one human touch. The annual pentest and the quarterly scan were already too slow for human attackers; they are not even in the same time zone as an agent that tries 148 things against Redis, gives up, and finds your actual zero-day in the next 11 minutes.
The only thing that keeps pace with an autonomous offensive agent is an autonomous defensive one that tests the same way. Zero Hunt runs an autonomous AI red team on-premise, on its own private models with a human in the loop for every action that matters: a 10-agent swarm that fingerprints the exact installed version, maps the live API, and tries to actually drive the vulnerable code path on your MONAI Label — or your Redis, or your Qdrant — in an ephemeral hardened container that never touches the host. It writes a fresh per-target exploit rather than pulling one from a database, which is exactly how you learn that PUT /datastore/image is reachable and writable before someone else's agent does. Because it runs continuously, the 48-hour window in which this campaign operated is not a window at all. See automated penetration testing for how continuous validation differs from the point-in-time test, and our earlier coverage of autonomous AI agents attacking live targets for where this is heading.
And when there is no released patch — as right now — the same engine's Remediation Advisor ranks the exposure by real exploitability (in-the-wild first, then reachability on your estate, then CVSS), hands you the exact compensating control to hold the line until the project ships a release, and re-verifies with the same proof that found the hole, so you know the path is closed and not just the version bumped.
Every vulnerability CISA lists as exploited, with federal due dates: CISA KEV tracker →
Patching one CVE at a time? A CTEM program validates which exposures are actually exploitable before the next advisory: what CTEM is →
Is this exploitable in your environment?
Zero Hunt answers that on your own network: an autonomous AI red team on an on-premise appliance, running on private AI, black-box or gray-box, with a human approving every step that matters. Proof of what is exploitable, the fix, and signed evidence — your data stays on your network.